Machine Learning-Based Prediction of Power Conversion Efficiency in ZnO Dye-Sensitized Solar Cells: Effects of Nanostructure and Synthesis Route

Data-driven methods provide effective tools for understanding dye-sensitized solar cell (DSSC) performance. This study extends a foundational machine-learning framework for ZnO-based DSSCs by adding 175 observations from literature published during 2021–2026, which expands the dataset from 298 to 473 experimental observations spanning 21 years. Six machine-learning models (Decision Tree, Random Forest, Support Vector Machine, K-Nearest Neighbors, XGBoost, and Artificial Neural Network) were compared using grid-search hyperparameter tuning, 5-fold cross-validation, and an 80:20 holdout split. XGBoost achieved R² = 0.59 on the modern seven-feature dataset, compared with R² = 0.48 for Random Forest on the legacy hydrothermal-only dataset. Initially, synthesis method accounted for 44.7–47.2% of feature importance, while nanostructure class contributed 36.6–37.9%. After synthesis method was separated into chemical synthesis route and film deposition technique, nanostructure class became the dominant predictor (49% ± 15%, 20-split average), and synthesis route and deposition technique each contributed about 12%. Mean power conversion efficiency (PCE) increased from 1.39% in 2005–2020 to 2.34% in 2021–2026 (Mann-Whitney U, p < 0.001; Cohen's d = 0.54), while maximum PCE increased from 7.66% to 9.86%. Cross-era validation on 17 modern hydrothermal observations from three studies indicated negative transfer for Random Forest (R² = −0.22), although three of six model families showed modest positive transfer. Given the small, low-variance test set dominated by a single study, this result is indicative rather than definitive. The findings point to the importance of technique diversity and careful feature definition when predicting the performance of ZnO-based DSSCs and when interpreting literature-derived experimental data.